Imberman 2011 - Achievement and Behavior in Charter Schools

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS: DRAWING A MORE COMPLETE PICTURE
Scott A. Imberman*
AbstractI use a long panel with broad grade coverage to establish whether charter schools affect cognitive and noncognitive skill formation. Schools that begin as charters generate large improvements in discipline and attendance but not test scores, with the exception of math in middle schools. This suggests improvements in noncognitive but not cognitive skills, although these improvements do not persist if students return to regular public schools. Charters that convert from regular public schools have little impact on either skill type. These results are robust to potential biases from selection off of precharter trends, attrition, and persistence.
I.
Introduction
HE charter school movement is one of the fastestgrowing education reforms in the United States today. Charter schools operate under a contract, called a charter, with a government agency. These schools are provided a degree of autonomy from local school boards and freedom from some regulations in return for additional accountability requirements. Despite often being managed by private organizations, charters are public schools and receive almost all of their funding from government sources. Since 1997 the number of charter schools in the United States has increased almost sixfold, and the number of charter students has more than doubled since 1999, as is shown in gure 1. As of 2006, 1.15 million students nationwide attended charter schools. One of the largest questions in the literature is how charter schools affect the outcomes of students who attend them. It is unclear whether charters are benecial or detrimental to students. On the one hand, charters have fewer regulatory burdens and are at higher risk of being shut down if they underperform. This provides incentives to increase effort. On the other hand, charters have high levels of student turnover, and eliminating certain regulations may be detrimental to students. In addition to this theoretical ambiguity,
Received for publication October 30, 2007. Revision accepted for publication November 12, 2009. * University of Houston I thank the Maryland Population Research Center for nancial support. I extend my sincerest gratitude to the employees and administrators of an anonymous school district for providing me with data and assistance and for making this project possible. I am especially grateful for the guidance and assistance provided by my dissertation advisor, Mark Duggan, and the immeasurable help from Steven Craig. In addition, I give special thanks to Judy Hellerstein, Bill Evans, Jeff Smith, two anonymous referees, and the editor, Michael Greenstone, for helpful suggestions and advice. I also thank Rajashri Chakrabarti, Ken Chay, Aimee Chin, Jose Galdo, Jonah Gelbach, Ginger Jin, Beom-Soo Kim, Melissa Kearney, Adriana Kugler, Jordan Matsudaira, Jennifer King Rice, John Rust, Seth Sanders, John Shea, Barbara Sianesi, Alex Whalley, Ye Zhang, Ron Zimmer, and seminar participants at Georgia State University, Stanford Institute for Economic Policy Research, Mathematica, RAND, University of Houston, University of Maryland, UNC-Chapel Hill, Urban Institute, Virginia Tech, APPAM, the North American Summer Meetings of the Econometric Society, and SEA. This work was done as part of my dissertation at the University of Maryland. All errors remain my own. The online appendix referred to throughout the article is available at http://www.mitpressjournals.org/doi/suppl/10.1162/REST_a_00077.
the empirical evidence has been mixed. Of the papers that use more advanced econometric techniques, some researchers nd insignicant or negative impacts of attending a charter school (Betts et al., 2006; Bifulco & Ladd, 2006; Buddin & Zimmer, 2005; Hanushek et al., 2007; Sass, 2006; Zimmer et al., 2008; Zimmer & Buddin, 2003, 2006), while others nd positive impacts (Booker et al., 2007; Clark, 2009; Hoxby & Murarka, 2009; Hoxby & Rockoff, 2004; McClure et al., 2005; Solmon & Goldschmidt, 2004; Solmon, Paark, & Garcia, 2001).1 Therefore, the effect of charter schools on student outcomes is unclear. One of the potential reasons for the wide variation in results is that some charter schools may not see test scores as their primary output. Many charters focus on students with special needs, such as those who are over age for their grade, new immigrants, or students who have difculty behaving well in a normal school environment. These students often need instruction not just in their academic abilitycognitive skillsbut in motivation, self-esteem, and self-discipline, which are types of noncognitive skills. The distinction between how cognitive and noncognitive skills are affected by education interventions has become increasingly important in the light of recent research showing that noncognitive skills have substantial inuence on labor market outcomes and degree attainment (Heckman, Stixrud, & Urzua, 2006; Heckman & Rubinstein, 2001). Thus, if charters improve noncognitive skills while having a limited effect on cognitive skills, they could still be effective tools for improving students outcomes in later life. Unfortunately, it is difcult to measure noncognitive skills directly. While testing of achievement is common in schools, other skills are rarely, if ever, tested. However, Heckman et al. (2006) establish that both cognitive and noncognitive skills improve behavioral outcomes. Hence, if charters improve behavior and attendance rates, that could be indicative of improvement in noncognitive skills even if cognitive skills measured by achievementdo not improve. In particular, attendance is a logical proxy for self-discipline and motivation, and using student behavior as a proxy for noncognitive skills has been done in Segal (2009a). In order to answer these questions, I use a unique data set from a large urban school district in the Southwest (LUSD-SW) to provide a broad look at achievement, student discipline, and attendance. Using these data, I establish whether there is evidence that charter schools affect cognitive or noncognitive skills. To my knowledge, no other study has
1 Most of these papers use xed effects or some similar identication strategy. Hoxby and Muraka (2009), Hoxby and Rockoff (2004), and McClure et al. (2005) instead use admission lotteries into charters as natural experiments.
The Review of Economics and Statistics, May 2011, 93(2): 416435 2011 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology
ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

Figure 1.Charter Growth In the United States
417
Sources: 19971998, U.S. Department of Education National Charter School Reports. 19992003: U.S. Department of Education Common Core of Data. 2005: National Alliance for Public Charter Schools. 2006: Center for Education Reform. 2004 data are unavailable, so a linear interpolation is provided.
considered how charters affect both of these skill sets. This panel is also useful in that it provides wider coverage over time and grades than any previous study of charter schools. In total, I have test score data for grades 1 through 11 over nine years, during which charter schools were operating in each year. Attendance and discipline data covers grades 1 through 12 over thirteen years. Using test scores with wide grade coverage appears particularly important, as limiting to the grade spans covered in Hanushek et al. (2007) and Booker et al. (2007) provides different results. In addition, I am able to look in detail at long-term impacts of charter schools while students are enrolled in charters and after they return to regular public schools. Both of these have rarely been studied in prior work and only on samples that did not include high school students.2 Whether charter schools generate lasting impacts is particularly important. For the foreseeable future, the stock of charter schools in the United States will be small relative to noncharters. As a consequence, most students who enter charters in elementary and middle school will return to noncharter schools before leaving the public school system. If charters provide short-term benets but no long-term benets, the usefulness of these schools for generating human capital will be limited. Hence, through studying multiple outcomes, using a broad base of students over a long time frame, and analyzing long-term effects, I provide a comprehensive and wide-ranging analysis of charter school impacts on charter students. I separate my analysis by two types of charters: start-ups and conversions. Previous work has shown these schools
2 Booker et al. (2007) nd that student performance improves as time in charters increases and also when they leave charters. Bifulco and Ladd (2006) also look at how students perform in their rst year in a charter and later years separately and nd that rst-year results were considerably worse. My strategy also differs from these two studies as I am able to instrument for students charter exit decision, which affects both the time in charter and persistence results.
differ in their impacts (Sass, 2006; Bifulco & Ladd, 2006; Zimmer and Buddin, 2003). As a result, considering them together could lead to aggregation bias.3 In addition, identifying whether these schools provide different impacts may have policy implications, since states and districts could allow only one type when starting a charter program, as is the case in Iowa, Mississippi, and Nevada.4 Start-up charters are schools with voluntary enrollment that begin as charters. Conversions in LUSD are schools that were previously regular public schools that convert to charter status. They keep the same staff, location, and attendance zones; thus, most of their students are assigned based on location of residence like any regular school. Both types of charters benet from exemptions from some regulations. Nonetheless, we would expect little impact from conversions since the change in the structure of the school is minimal. My results show this generally to be true. Start-up charters, however, generate impacts on student outcomes. While I nd no statistically signicant effect overall from attending a start-up charter on test scores, there is an improvement in math for students in middle school grades of 0.07 to 0.18 standard deviations. Nonetheless, despite these test score ndings, students garner large and statistically signicant improvements in attendance and discipline from attending start-up charters. On average, attendance rates increase by 2.4 percentage points, or 23% of the absence rate in the year prior to charter entry. Start-ups also reduce annual disciplinary infractions by 0.5 to 0.8 instances, a very large impact relative to the average of 1.1 infractions in the year prior to entry. However, these results do not persist after
3 Hanushek et al. (2007) and Booker et al. (2007) do not disaggregate by these charter types. However, in Texas, only a few charters are conversions. 4 According to the Center for Education Reform, Iowa and Mississippi permit only conversions. Nevada permits only start-ups. The other 37 states that permit charters and Washington, D.C., allow both.
418
THE REVIEW OF ECONOMICS AND STATISTICS
students return to noncharter schools. In particular, attendance and discipline impacts disappear immediately after students return to regular public schools. Therefore, while the impact of start-up charters on cognitive skills is small, they appear to generate a substantial improvement in noncognitive skills; however, the drop-off after students return to regular public schools shows that these skills require continual reinforcement in students. Nonetheless, one should be cautious in this interpretation as there are other potential explanations for the behavior results. One possibility is that charters differ from regular public schools in how they enforce or report discipline. This is an important concern, and although it is not possible to completely rule out this explanation, I provide a series of tests and arguments that show that it is unlikely this is driving the results. Of particular importance is that the strong attendance results, which are much harder for the charters to manipulate than discipline, serve to reinforce the discipline results, and the fact that both measures show large and significant improvements provides evidence of noncognitive skill improvements. Another possibility is that the results may reect institutional differences between charters and noncharters rather than skill formation. The results for persistence are consistent with this theory. In addition, while Heckman et al. (2006) show that noncognitive skills themselves improve labor market outcomes and Segal (2009b) establishes a link between teacher-reported student behavior and later-life earnings, a causal impact of behavior and school attendance on wages has yet to be established. Nevertheless, even if the discipline and attendance impacts do not translate into improvements later in life, they are potentially important outcomes in their own right since parents consider behavior to be an important factor in the decision on whether to send their children to charters. In a survey of Texas charter parents, Weiher and Tedin (2002) show that only one-quarter list test scores as the primary reason for sending their children to charters, while more than two-thirds cite moral values, discipline, or safety. Thus, if the student exhibits improvements in behavior and increased school attendance, parents utility would likely increase even if those improvements are not permanent. In addition to having multiple outcomes and the ability to assess long-term charter impacts, I also address some econometric issues. One potential econometric problem is that the assumptions underlying xed effects are invalid if students choose to attend charter schools based on changes in outcomes. If this occurs, then the estimates of charter impacts may be contaminated by mean reversion. This phenomenon has been widely noted in the job-training literature (Ashenfelter, 1978; Heckman & Smith, 1999) while in education, mean reversion has been shown to occur in standardized exams (Chay, McEwan, & Urquiola, 2005). Hanushek et al. (2007) use interrupted panel estimates to argue that this selection does not pose a problem. While I nd some graphical evidence of this type of selection in start-up charters, my
interrupted panel estimates show little change from the main regression estimates. Consequently, although this selection does appear to exist, it does not substantially affect the impact estimates. Another potential problem is nonrandom attrition. This could create bias if the charter students leave the district at a rate that differs from that of public school students. While LUSD is a central city school district, it is bordered by ten school districts. It also has many state charter schools and private schools within its boundaries. As such, there are a lot of educational options for parents, so the attrition bias is a concern. Other work has also shown that attrition bias in charter research is a potential problem. For example, Hanushek et al. (2007) nd that charter students leave Texas public schools at more than 2.5 times the rate of noncharter students. To address this, in addition to tests for differential attrition, I use a unique semiparametric attrition adjustment procedure for xed-effects analyses proposed by Kyriazidou (1997). This procedure has not previously been used to assess charter impacts. These results suggest that nonrandom attrition does not have a substantial effect on the charter impact estimates.
II. LUSD Characteristics and Data Description
In this paper, I use a unique panel of student-level administrative records from a large urban school district in the Southwest. LUSD-SW was one of the rst school districts in the United States to institute a charter program. The program, which began in 1996 with 2 schools, expanded to half its current size in 1997 and 1998. By 2006 there were 32 charter schools, 23 start-ups and 9 conversions.5 Students from nearly 300 noncharter schools are also observed in the data. During the time period studied, approximately 50 state charter schools were also in operation for which I do not have data. Figure 2 shows the evolution of the charter program in LUSD by examining the fraction of enrollment in start-up and conversion charters. In 1997 and 1998, all of the conversions obtained charter status, after which their enrollment shrank relative to total growth. Most of the start-up charters opened in 1998 and 2001, but their population steadily increased over time. As of the 20062007 school year, 4% of students in LUSD attended a charter school. Table 1 provides some summary information about charter schools and students. Start-up students are more likely to be minority, poorer, and more at risk than noncharter students. The schools are smaller on average and spend more per student on instruction but less on other expenses. Conversions are also more
5 One charter existed under contract with LUSD prior to the enactment of the states charter laws and then promptly switched. Since enrollment is voluntary, I dene it as a start-up. Some start-ups reside on the campus of an existing school but are considered independent schools and have voluntary enrollment. One conversion charter maintains a large gifted and talented magnet program. In order to prevent the impact of this program from inuencing the charter impact estimates, I drop any student who attends that school from the analysis, leaving eight conversions in the nal sample.

Figure 2.LUSD-SW Charter Enrollment by Year, Grades 112
419
heavily minority and poorer than regular public schools but are otherwise similar.6 The LUSD-SW data set includes information on scores from the Stanford Achievement Test, disciplinary records, attendance rates, and a number of student characteristics. By combining these outcomes I assess to what extent charter schools affect both cognitive and noncognitive skill development. The data cover the 199495 to 20062007 academic years, and I am able to follow individual students for as long as they attend school in LUSD, providing a long time series on many students.7 Hence, I am also able to look at how these skills develop in the long term both during and after charter exposure. Stanford Achievement Test scores are available starting with the 199899 school year, providing nine years of test score data. The exam is norm referenced, so it reects achievement relative to a national sample of students. I analyze three exam subjectsmath, reading, and language given in grades 1 through 11. For each subject, I standardize the scaled scores to be mean 0 and standard deviation 1 within grade and year across LUSD. Hence, impacts are measured in standard deviation units relative to the district average. In order to ensure that the impacts are analyzed on the same sets of students across all tests, after standardizing the scores, I limit the sample to students who have scores for all three
6 Start-up charters tend to be spread across all grade levels. There are thirteen start-ups covering at least one elementary grade, twelve covering at least one middle school grade, and eight covering at least one high school grade. Since charters often diverge from the standard grade structure, these numbers refer to only 23 schools. Seven of the eight conversions I study cover at least one elementary grade. Three cover at least one middle school grade, but only one of these includes grades 7 or 8. There are no high school conversions. 7 After dropping observations for early education, prekindergarten, and kindergarten, 56% of students who are rst observed in the data prior to ninth grade have at least four observations. In addition, only 27% of startup charter and 21% of conversion charter students have neither pre- nor post-charter observations.
exams.8 The nal sample for test scores contains approximately 1.14 million student-year observations, including 15,000 start-up and 20,000 conversion charter observations. In table 1, I show that for both types of charters, the test score means are not statistically signicantly different from those of regular public schools. Discipline and attendance records of students cover 1994 1995 through 20062007 for all students in grades 1 to 12. The attendance rate provides the percentage of days the student attends school while enrolled. Discipline records provide information on the type and length of punishment for any infraction that results in an in-school suspension or more severe punishment.9 I use the total number of disciplinary infractions per year in most of the analyses. Across LUSD, 17% of student-year observations have at least one infraction, and 9% have multiple infractions. For start-up charters, those gures are 8% and 3%, respectively, and for conversions, they are 14% and 6%. In total, the sample for discipline and attendance contains 2.23 million observations, of which 20,000 are students in start-ups and 40,000 are students in conversions. Table 1 shows that on average, start-up charters
8 Some students who are not procient enough in English in grades 18 took a separate Spanish language exam called Aprenda. While I have data on these exam results, the scores are not directly comparable to those of students taking the English exam so I do not include them in the analysis. Almost all students who take Aprenda are in grades 15 and account for 24% of all test takers in those grades. One concern is that since start-ups include fewer limited English procient (LEP) or special education students, and hence only a handful of Aprenda takers, the Stanford Achievement Test results may be biased. However, test score regressions limited to students who are not classied as LEP or special education for the duration of the test sample show similar results to the baseline regressions. Hence, this is not a substantial problem. 9 Unfortunately, infractions that generate less harsh punishments, such as detention, are not in the data. In addition, the records provide only limited information on the type of infraction, since 80% of infractions are unspecied student code violations. Nonetheless, a few severe infractions such as substance abuse, criminal behavior, and, for 20022003 and later, ghting are identied in the data.
420

Table 1.Summary Statistics by Charter Status Demographics (Student-Level Observations) Outcomes (Student-Level Observations) Conversion 0.48 (0.50) [1.1] 0.01 (0.11) [7.8] 0.50 (0.50) [1.2] 0.48 (0.50) [0.5] 3.74 (2.07) [3.0] 0.79 (0.41) [9.4] 0.08 (0.27) [1.6] 0.07 (0.25) [1.2] 0.29 (0.46) [0.5] 0.59 (0.49) [0.2] 0.11 (0.32) [0.1] 0.05 (0.23) [3.6] 0.09 (0.29) [1.7] 3,688 (743) [1.7] 455 (158) [1.0] 918 (389) [0.1] 599 (277) [1.7] Variable Math Reading Language Disciplinary infractions Substance abuse infractions Violent crime infractions Fighting infractions Attendance rate Noncharter 0.03 (0.97) 0.03 (0.98) 0.02 (0.98) 0.44 (1.37) 0.0074 (0.13) 0.0047 (0.09) 0.0359 (0.22) 94.4 (8.89) Start-Up 0.02 (0.93) [0.1] 0.06 (0.90) [0.5] 0.04 (0.91) [0.2] 0.16 (0.73) [4.3] 0.0073 (0.12) [0.0] 0.0024 (0.07) [1.4] 0.0163 (0.15) [2.1] 93.2 (11.98) [0.6] Conversion 0.11 (0.95) [1.0] 0.11 (0.92) [0.9] 0.08 (0.94) [0.6] 0.35 (1.29) [0.4] 0.0007 (0.03) [5.6] 0.0019 (0.06) [2.7] 0.0498 (0.27) [0.7] 95.7 (5.97) [1.5]
Variable Female
Noncharter 0.49 (0.50)
Start-Up
0.47 (0.50) [0.7] White 0.10 0.03 (0.30) (0.16) [4.5] Black 0.33 0.31 (0.47) (0.46) [0.3] Hispanic 0.54 0.66 (0.50) (0.47) [1.5] Grade level 5.85 6.53 (3.33) (3.50) [0.5] Free lunch 0.60 0.58 (0.49) (0.49) [0.5] Reduced-price 0.07 0.11 lunch (0.25) (0.31) [6.1] Other economic 0.05 0.11 disadvantage (0.23) (0.32) [2.9] Limited English 0.25 0.18 prociency (0.43) (0.39) [1.1] At-risk status 0.58 0.65 (0.49) (0.48) [0.9] Special education 0.11 0.05 (0.32) (0.21) [6.0] Gifted and 0.10 0.01 talented (0.30) (0.12) [6.3] Immigrant 0.13 0.13 (0.33) (0.34) [0.2] School characteristics (School-level observations) Instructional 4,054 5,511 spending per (5,668) (5,341) student [1.8] Adminstrative 519 164 spending per (1,670) (350) student [4.6] Other spending 936 254 per student (1,363) (322) [8.1] Enrollment 739 225 (509) (263) [7.9]
Standard deviations in parentheses. Absolute T -statistics from a regression of the variable on start-up and conversion status in brackets. Test scores are standard deviation units from scale scores normalized within grade and year. Spending gures are per pupil. Demographics, attendance, and general infractions include 2.45 million student-year observations. Due to limited years of availability substance abuse and violence infractions include 1.79 million observations, while ghting includes approximately 700,000 observations. Test scores include 1.2 million observations. School characteristics include approximately 3,900 school-year observations.
have signicantly lower overall infraction rates and fewer ghting infractions than noncharters. Conversions have fewer substance abuse and violent crime infractions. Means for attendance are not statistically signicantly different from regular public schools for either charter type. Thus, the summary statistics provide suggestive evidence of there being little impact on test scores for either charter type but some improvement in behavior for start-up charters. This is consistent with charters improving noncognitive skills but not
cognitive skills. To conrm this result, I turn to regression analysis.

III. A. Baseline Model Empirical Strategy
In order to identify whether charters improve cognitive and noncognitive skills, I use the following individual xedeffects model:
421
yit = + C Conversionit + S Startupit + Demogit + Switchit + Gradeyearit + i + it , (1) where yit is some outcome measure for student i at time t, Conversionit and Startupit are indicators for the type of charter the student is enrolled in Demogit is a vector of timevariant observable demographic characteristics, Gradeyearit is a set of grade-by-year indicator variables that account for changes in outcomes over time and grade level, i is an individual xed effect, and it is an independent and identically distributed (i.i.d.) error.10 Switchit is a set of variables that dene whether a student changes schools in year t due to a structural change from normal grade progression, a nonstructural change for some other reason, or if it is the students rst year in the district. All of these variables are interacted with indicators for the students grade level.11 A problem with this strategy is that prior test scores play a role in current achievement. To address this, researchers often use a value-added (or gains) version of the xedeffects model where the dependent variable is the annual change in outcomes.12 This is equivalent to assuming that if lagged achievement is an explanatory variable, then its coefcient equals 1. Since the role of lagged achievement likely decays, it may be preferable to have a model that explicitly includes lagged achievement as an explanatory variable, as in Hanushek et al. (2007) and Sass (2006): yit = + yi,t1 + C Conversionit + S Startupit + Demogit + Switchit + Gradeyearit + i +
13 it .
suggested that factors in childrens distant youth play important roles in later achievement (Todd & Wolpin, 2007), suggesting that twice-lagged scores are unlikely to be exogenous, and thus the estimates in these papers may be biased. As an alternative, I use both levels and value-added models. In expectation, these two models bound the true estimate, which would be identied by equation (2). I provide the proof of this in the appendix.14 Therefore, I am able to identify charter impacts within a range of values while avoiding biases that are introduced by endogenous instruments.
B. Potential Biases from Selection
(2) Since lagged scores are endogenous, these papers instrument with twice-lagged scores. However, recent research has
10 Demog includes free-lunch status, reduced-price lunch status, other it economic disadvantage, whether the student is a recent immigrant, and whether a parent is a migrant worker. Other economic disadvantage indicates the student does not receive free or reduced-price lunch but does receive other poverty assistance. More detailed denitions are in the online appendix. 11 Hanushek et al. (2007), Booker et al. (2007), and Bifulco and Ladd (2006) suggest that properly controlling for student switches is important for separating charter impacts from the effects of switching schools. I follow Bifulco and Ladd and dene a nonstructural switch as switching into a school where less than 10% of a students previous class switches into the same school. Conversely, a student undergoes a structural switch when more than 10% of his or her previous class switches into the same school. Twelve percent of student-years undergo nonstructural switches, 11% of student-years undergo structural switches, and 12% are students moving from outside the district. 12 Alternatively once could use a random trend model where both sides of the estimating equation are differenced and then demeaned. While this allows for individual time trends, it substantially reduces precision and could exacerbate bias if the trends are nonlinear. Nonetheless, estimates from this model are similar to the baseline levels models. 13 Sass (2006) uses an Arellano-Bond model. I conducted regressions using this procedure and found results that are similar to my baseline levels model. Hanushek et al. (2007) rst-difference their data to remove the xed effects rather than demean and then instrument the change in lagged achievement with the twice-lagged level.
While attending any school is a choice, parents are not restricted by having to live in a specic attendance zone or meet some transfer qualication for most charters. As a result, selection into charters may be a more substantial problem than selection into regular public schools. The student xed-effects analysis used in this paper corrects for selection into charters based on unobserved characteristics that do not change over time, such as innate ability. Nonetheless, one may be concerned of selection due to time-varying factors. While I cannot eliminate all types of selection that could play a role, I am able to check for a few specic types that would be particularly important in the charter context.15 My main ndings, that there is little improvement in cognitive skills but substantial improvement in noncognitive skills from attending a start-up charter, are robust to these potential biases. The rst issue I consider is whether entry into charter schools is based on pre-entry trends in the dependent variable (Booker et al., 2007; Hanushek et al., 2007; Bifulco & Ladd, 2006; Sass, 2006). It is possible that students enter charter schools due to changes in test scores or behavior or a change in some strong correlate with these outcomes. Such a situation has been widely noted in the job-training literature and is commonly called Ashenfelters dip (Ashenfelter, 1978; Heckman & Smith, 1999). Since a parent may see a drop in performance as an indicator that the current school does not meet his or her childs needs, it is reasonable to believe that students change schooling environments in response to poor
14 This framework is similar to that proposed by Guryan (2001), which uses difference-in-differences and lagged-dependent variable models to bound a true estimate where it is not clear whether selection is based on xed characteristics or previous outcomes. 15 Another solution is to use oversubscription lotteries as natural experiments (Hoxby & Murarka, 2009; McClure et al., 2005; Hoxby & Rockoff, 2004). While a lottery-based strategy has substantial advantages over xed effects, there are three potentially undesirable aspects. First, oversubscribed schools are likely of higher quality than schools with spaces available; hence, a comparison of lottery winners and losers will identify the impacts for only the best charter schools. Second, lottery studies may be subject to more attrition bias than panel studies, since parents who lose lotteries may be more likely to send their children to private schools or other districts than those who win. Third, often lotteries are limited to a very small number of schoolsfour in the case of Hoxby and Rockoff (2004) and one in the case of McClure et al. (2006). Nonetheless, Hoxby and Muraka (2009) are able to address all three of these concerns well, as their sample of New York City charter schools has forty schools with lotteries and they appear to have little attrition bias.
422

Figure 3.Transitions between School Types, 19982005, Grades 111
performance. This violates the strict exogeneity assumption for xed effects, which requires that future and past outcomes are not affected by current charter status. While much of this problem is dealt with by appropriately controlling for students switching status, I provide evidence that some small dips remain for start-up charters. Hence, I follow Hanushek et al. (2007) in the use of interrupted panel estimates to check for bias from endogenous charter entry. In addition, regressions that look at how charter impacts vary with time spent in a charter are used to check for mean reversion, which would be a sign that the estimates suffer from this bias. Another problem is that some parents may choose to leave LUSD altogether if their student performs poorly in the charter school. If this occurs at a different rate from that in regular public schools conditioning on observables and student xed effects, then there are time periods when these students should be observed but are not. This could lead to attrition bias. For LUSD, this is particularly important because of the many other options available to students, including private schools, state charters, and more than ten suburban school districts. Figure 3 provides some suggestive evidence that endogenous attrition is a potential problem. While about 16% of noncharter students exit LUSD each year, 26% of start-up charter students attrit. Though not shown in the gure, the differences are more dramatic over longer time periods. For example, 39% of noncharter third graders between 1998 and 2000 are no longer in LUSD ve years later, while that number is 43% for conversion students and 63% for start-up students. These statistics may simply reect different characteristics of the schools, such as different grade levels covered or the types of students who attend. Indeed, regressions of attrition propensity on charter status, including all of the covariates in equation (1), show no statistically signicant relationship between charter status and attrition propensity
at all grade levels.16 However, a model that interacts charter status with test scores shows that start-up students with higher reading scores and conversion students with higher math scores are less likely to attrit than comparable noncharter students. These results can be found in the online appendix. While differential attrition appears to be only a minor issue, the evidence is not denitive. As such, I use a nonparametric procedure for correcting sample selection in individual xed-effects models proposed by Kyriazidou (1997) to further check for attrition bias. Her insight is that since xed effects correct for attrition based on time-invariant factors, one can correct for endogenous attrition due to time-variant factors by weighting toward observations where there is no change in attrition propensity. To produce Kyriazidous (1997) estimator, one must rst dene the selection equation, sit = Wit + i + it , (3)
where sit equals 1 if the student is no longer in the sample and 0 otherwise, Wit is a set of variables that are observed whether or not the individual has attrited, i is an individual specic xed effect, and it is random i.i.d. error. Wit need not contain all (or any) of the variables in the outcome equation, but it does need to contain an exclusion restriction. In this paper, I use a model that includes the students last observed start-up or conversion charter status, free lunch status interacted with grade level, reduced-price lunch interacted with grade level, other economic disadvantage interacted with grade level, recent immigration status, whether a parent is a
16 In the model with a pooled estimate across all grade levels, the coefcient on attrition from start-ups is 0.037 (s.e. 0.025) and conversions is 0.003 (s.e. 0.010).
423
migrant worker, grade-by-year indicators, and whether the student is ineligible to attend his or her previous school due to his or her predicted grade not being offered, which serves as the exclusion restriction.17 The assumption underlying this restriction is that being in a grade beyond the students last observed schools highest grade is correlated with attrition but not student outcomes.18 While one may be concerned that these students are invariably older and thus this may correlate with outcomes, particularly discipline and attendance, the inclusion of grade-by-year xed effects mitigates this. The model is estimated using a conditional xed-effects logit. The outcome equation in my model is the baseline charter impact equation. After removing the xed effect in the outcome equation through rst-differencing, Kyriazidou (1997) argues that in observations where (Wit Wis ) = 0 for s < t, the individual has not had a change in circumstances that affects attrition. Since a students innate tendency to switch schools is captured by xed effects, I can generate consistent estimates of , the charter effect, by using only those observations where this holds true. Since limiting to observations where (Wit Wis ) = 0 would reduce power substantially, Kyriazidou (1997) proposes using kernel weights to focus the analysis on observations that are close to (Wit Wis ) = 0. Therefore, in the second stage, I run a rst-differenced version of equation (1) weighted by it,n = 1 K hn (Wi,t Wi,t1 ) , hn (4)
toward 0.20 To solve both problems, I conduct analyses that include indicators for whether a student is in a post-charter period, which allows us to compare charter impacts directly to pre-charter periods while identifying whether charter impacts persist after students return to regular public schools:21
4
yit = + S,0 Startupit +

k=1
S,k PostStartupk
4
+ C,0 Conversionit +
k=1
C,k PostConversionk
+ Demogit + Switchit + Gradeyearit + i + it , (5) where PostStartupk and PostConversionk denote the students being in year k after leaving the charter. For k = 4, I include any time period after the third year. To account for students endogenously exiting charters, I instrument for being in a post-charter period with whether the student has the listed number of predicted grades past his or her previous charters highest grade covered. Predicted grades are calculated using normal grade progression starting from the year of charter entry.22 This avoids potential complications from the charters retention policies. For example, the estimator for being one year after start-up is instrumented using whether the student is one predicted grade level higher than the highest grade covered in the charter school he or she attended. Hence, I am able to isolate the effects from students who are forced out of charters due to being grade ineligible instead of those who leave voluntarily, possibly because they do not perform well in the charter.
IV. Results
where K is a kernel function with bandwidth hn and (Wit Wi,t1 ) is the rst-differenced linear prediction from the selection model estimation.19 A third problem that could arise is bias from students leaving the charters and returning to the regular public schools. Both the possibility that students leave because they perform poorly in charters and that charters have long-term impacts on outcomes can contribute to the bias. In the rst case, if students leave charters prematurely due to poor performance, then they will reduce the number of charter period observations and the inuence of bad charters. In the second case, charters that have long-term impacts that persist when students return to regular public schools will also affect periods when students are not in charters and bias impacts
17 For attrited observations, I use the grade the student would have been in assuming normal grade progression. 18 The idea behind this exclusion restriction is that a student would be more likely to leave the district if she has to switch schools anyway; that is, the relative costs of leaving the district fall if students are forced to switch schools. 19 The appropriate bandwidth is found using the mean squared error (MSE) minimization procedure described in Kyriazidou (1997). Since the MSE minimizing bandwidth is sensitive to the initial value, I follow Dustmann and Rochina-Barrachina (2000) in using the initial value that is as close as possible to the constant for the MSE minimizing bandwidth as asymptotic theory says the two should converge. I use the gaussian kernel to generate the weights.
A. Charter School Impacts on Student Outcomes
I now turn to my main results. First, I consider test scores to see how charter schools affect cognitive skill development. In gure 4, I provide graphs that trace out the residuals of student outcomes from a xed-effects regression, including all of the covariates in model (1) except charter status. This allows me to look at how outcomes change as students enter charters net of xed student characteristics, economic status, immigration status, grade level, and time. The top row shows start-up charters, which are the type of charter in which one would most expect to see charter impacts. While math scores seem to increase after entry, reading and language show far less improvement. The graphs also show some evidence of
20 This is a violation of the strict exogeneity assumption for xed effects, which implies that outcomes in future and past periods cannot be correlated with current charter status conditioning on observables and the xed effect. 21 It is possible to be in a conversion charter and also in a post-start-up period at the same time (and conversely for start-up and postconversion); however, only 2% of charter students ever attend both types. 22 While I do not include kindergarten in the analysis, as there are no test scores for this grade and enrollment is optional in LUSD, I am able to identify whether a student enters a charter in kindergarten.
424

Figure 4.Test Scores before and after Charter Entry
selection into the charter off precharter trends since test scores drop in the year immediately prior to entry (1), which could potentially bias the estimates. For conversions, test scores appear to fall off slightly after entry. The graphs suggest that with the possible exception of math in start-ups, cognitive skill improvement is small at best. Table 2 provides the main test score results for this paper, along with the interrupted panel and Kyriazidou (1997) estimates. The standard errors for each regression are robust to heteroskedasticity and clustered by school. Although I focus on models that consider both conversion and start-up charters, it is useful to start by looking at a model where start-ups and conversions are constrained to have the same coefcient. In this case, there is a statistically signicant effect on math in levels of 0.07 standard deviations, but no statistically signicant impacts for any other estimate. When I disaggregate the estimate, the results for start-ups conrm the graphical analysis as there is no statistically signicant impact of startup charters on test scores in levels or value-added models.23 With the exception of math in levels, all of the point estimates are relatively close to 0. Consequently, these results suggest little improvement in cognitive skills in start-up charters on average. For conversions while math in levels is statistically signicant at 0.07 standard deviations, the value-added estimate is insignicant and thus I cannot say that conversions
have a math impact.24 In addition, this impact will disappear when I consider the role of persistence. A concern with using xed effects to study charter impacts is that since many students enter charters in kindergarten, by excluding earlier years the estimates are identied off students who may disproportionately benet from charters. I test the extent of this problem by conducting regressions restricted to students in grades 3 to 11, 3 to 8, and 4 to 8.25 The results for these are provided in the online appendix. If the xed-effects estimates in prior research are biased from excluding early charter students, then adding rst and second graders should change the estimates.26 When restricting grade spans to grades 3 to 11, there is little change from the baseline models in the test score impact estimates for startups or conversions. Hence, it appears that excluding grades 1 and 2 from the analysis does not bias the results. However, when I restrict to grades 3 to 8 or 4 to 8, the estimates for start-ups are statistically signicant and positive in both levels and value-added models for math and language and in
24 At-risk, LEP, gifted and talented, and special education status could be useful covariates, but they are also potentially inuenced by school quality, or charters may not classify students the same way as noncharters. As such, these variables would also be inappropriate as outcomes. Nonetheless adding LEP, at-risk status, gifted and talented status, and special education status to the baseline model has very little effect on the estimates. 25 Hanushek et al. (2007) and Booker et al. (2007) use, respectively, the grade spans 38 and 48. 26 One complication with this strategy is that, as noted by Hoxby and Murarka (2009), the benet of adding these grades to analyses may be small since test scores in these grades are less accurate measures of achievement than tests in later grades.
23 Levels models restricted to the value-added sample provide estimates that are similar for all outcomes and are available in the online appendix.

Table 2.Effect of Attending a Charter School on Test Scores. Math (1) OLS with student Baseline xed effects OLS with student Baseline xed effects Interrupted panel Drop year prior to charter Drop 2 years prior to charter Kyriazidou (1997) First differences attrition model (unweighted) MSE minimizing bandwidth Any charter Start-up Conversion Start-up Conversion Start-up Conversion Start-up Conversion Start-up Conversion 1/4 MSE minimizing Start-up bandwidth Conversion 0.071 (0.031) 0.072 (0.052) 0.070 (0.033) 0.060 (0.055) 0.081 (0.039) 0.056 (0.054) 0.084 (0.041) 0.049 (0.040) 0.073 (0.022) 0.053 (0.039) 0.071 (0.022) 0.065 (0.042) 0.044 (0.023) A: Levels Reading Language Number of (2) (3) Observations 0.023 (0.023) 0.015 (0.038) 0.029 (0.026) 0.010 (0.041) 0.040 (0.032) 0.009 (0.041) 0.042 (0.033) 0.003 (0.028) 0.029 (0.022) 0.004 (0.028) 0.028 (0.022) 0.005 (0.026) 0.018 (0.023) 0.033 (0.020) 0.022 (0.033) 0.041 (0.024) 0.013 (0.036) 0.048 (0.028) 0.006 (0.036) 0.043 (0.029) 0.011 (0.022) 0.054 (0.017) 0.012 (0.022) 0.054 (0.017) 0.018 (0.024) 0.063 (0.021) 1,141,480 1,141,480 Math (4) 0.026 (0.022) 0.026 (0.025) 0.026 (0.034) 0.062 (0.051) 0.070 (0.041) 0.041 (0.051) 0.070 (0.042) 0.059 (0.036) 0.063 (0.043) 0.075 (0.036) 0.051 (0.041) 0.100 (0.046) 0.015 (0.044)
425
B: Value-Added Reading Language Number of (5) (6) Observations 0.025 (0.016) 0.033 (0.024) 0.019 (0.021) 0.002 (0.038) 0.030 (0.035) 0.008 (0.038) 0.029 (0.036) 0.009 (0.033) 0.001 (0.028) 0.005 (0.034) 0.008 (0.028) 0.014 (0.032) 0.037 (0.037) 0.004 (0.016) 0.007 (0.023) 0.014 (0.021) 0.015 (0.031) 0.042 (0.027) 0.001 (0.032) 0.038 (0.029) 0.006 (0.024) 0.028 (0.022) 0.007 (0.026) 0.031 (0.021) 0.043 (0.037) 0.066 (0.031) 779,343 779,343
1,135,119
772,568
1,131,123
767,916
618,050
526,203
618,050
526,203
618,050
526,203
All regressions include an individual xed effect, free lunch status, reduced-price lunch status, other economic disadvantage, recent immigration status, whether a parent is a migrant worker, and grade-by-year indicators. Standard errors clustered by school in parentheses. , , denote signicance at the 10%, 5%, and 1% levels, respectively. Levels models cover 19982006 and grades 111. Value-added models cover 19992006 and grades 211. Because students are dened to be in the sample in the rst grade and year of each sample, these observations are dropped from the Kyriazidou attrition analysis.
levels for reading. This is indicative of heterogeneous effects across grade levels, an issue I explore further below. The next two rows show the interrupted panel estimates.27 These estimates account for the preentry dips seen in gure 4. These estimates are close to the baseline results regardless of whether one or two years prior are dropped. A possible exception is math for conversions in the value-added models where the impact increases to 0.07 standard deviations but is signicant only at the 10% level. The last three rows provide estimates using the Kyriazidou (1997) attrition correction procedure. Since this relies on a rst-differenced estimator rather than a xed-effects estimator, I use these to examine how much estimates change when the attrition corrections are applied rather than as true impact estimates. The unweighted model is the rst-differences corollary to the baseline model in the rst row. If attrition is a problem, then one would expect there to be large changes in the estimates as the weights are added. I show both models using mean-squared error minimizing weights and weights where the bandwidths are one-fourth of the MSE minimizing bandwidths. While start-up math impacts increase and become
27 In value-added models, Hanushek et al. (2007) keep the gain in the rst charter year as the difference between year t and t 1. While this reduces the bias from the precharter gain measures, it does not reduce bias from the excessive gain in the rst charter year. I modify the procedure such that the dependent variable in the value-added models is the average gain over the dropped years. That is, when year t 1 is dropped, I use the difference in test scores in t and t 2 divided by 2. Both strategies provide similar results for all outcomes.
statistically signicant in value-added models, levels models are still statistically insignicant and change little. There is also some increase for language in conversions. Nonetheless, these results at worst suggest that I may be slightly underestimating math scores for start-ups and language scores for conversions, but in general the baseline results hold up well. As such, I am unable to detect a statistically signicant effect of start-up charter schools on measures of cognitive skills. However, it is still unclear how charters affect noncognitive skills. Since many start-up charters in LUSD target students with behavioral problems and those who are classied as at risk of dropping out of school, these schools may focus more on noncognitive than cognitive skills. While some cognitive skill improvement spills into discipline and attendance, the lack of test score impacts implies that improvements in these outcomes would likely be due to noncognitive skill enhancement. Figure 5 provides the same graphs for discipline and attendance as gure 4 provides for test scores. While there is little change in either of these measures after charter entry for conversions, the gures show drops in disciplinary infractions and increases in attendance rates immediately after entry. In both cases, there is some reversion after the rst year, but both outcomes remain at levels that improve on preentry performance. As in the case with test scores, however, there is evidence of selection into charters off worsening attendance, so once again I need to use interrupted panel estimates to address this potential bias.
426

Figure 5.Discipline and Attendance, Before and After Charter Entry
Panel A of table 3 shows the baseline, interrupted panel, and attrition-adjusted results for discipline and attendance. I begin by providing results for an overall charter effect. The results show statistically signicant improvements in discipline of 0.2 to 0.4 infractions for charter attendees, while attendance improves by 1 percentage point. However, the combined estimate hides substantial differences between start-up and conversion charters. While there are no statistically signicant estimates in any model for conversion charters, start-up charters have large and statistically significant improvements in discipline and attendance. Using the levels and value-added model estimates as bounds, baseline estimates in the rst row show annual disciplinary infractions falling by between 0.5 and 0.8 incidences when students attend start-up charters. These estimates are statistically signicant at the 1% level. Average infractions in the year prior to charter entry are 1.1, so the start-up charter impact is between 45% and 73% of the precharter mean. Start-ups also have statistically signicant impacts at the 1% level on attendance. Baseline estimates show an increase of 2.3 percentage points. In the year prior to entry, attendance rates average 91.0%, so the attendance impact accounts for 26% of the precharter absence rate.28 In addition, despite the graphical evidence for the Ashenfelter dips, discipline and attendance results hold in
the interrupted panel and attrition-adjusted estimates. Therefore, it appears that start-up charters improve noncognitive skill formation. Since LUSD audits attendance by checking teachers logs with reported attendance, any systematic misreporting would require participation of administrators and teachers, which would be very difcult. Therefore, we can be condent that the attendance results reect actual behavioral improvements. Nonetheless, it is possible that the discipline results are due to differences in enforcement or reporting rather than actual behavior improvements. To address this concern, I provide a few pieces of evidence that at least part of the impact reects behavioral improvements. First, the large impact from attendance reinforces the discipline results since they are highly correlated and both reect noncognitive skill improvement.29 Second, regressions that use severe infractionssubstance abuse, violent criminal activity, and ghtingas outcomes show statistically signicant drops from attending a start-up charter. In levels models, substance abuse infractions fall by 0.015 (s.e. 0.006), violent crimes by 0.008 (s.e. 0.002), and
29 An OLS regression of attendance rate on the number of infractions, free-lunch, reduced-price lunch, other economic disadvantage, recent immigrant, parents migrant status, gender, race, and grade-year interactions provides a point estimate on infractions of 1.12 and a standard error (clustered by school) of 0.05. It is possible that higher attendance could generate more disciplinary problems in charter schools, as the marginal attendees would tend to misbehave. Nonetheless, this would cause me to underestimate the impact of charters on discipline. Since I nd large impacts on both attendance and discipline, it remains most likely that both result from a common underlying skill enhancement.
28 Regressions limiting the discipline and attendance analysis to students who took all three Stanford exams are similar for discipline. Attendance impacts are lower but remain statistically signicant at the 1% level. Results are provided in the online appendix.

Table 3.Effect of Attending a Charter School on Discipline, Attendance, and Retention Disciplinary Infractions (1) OLS with student xed effects OLS with student xed effects Interrupted panel Baseline Baseline Any charter Levels Attendance (2) Number of Observations Disciplinary Infractions (3) 0.233 (0.109) 0.498 (0.151) 0.005 (0.058) 0.789 (0.085) 0.024 (0.088) 0.797 (0.084) 0.024 (0.089) 0.849 (0.208) 0.034 (0.094) 0.842 (0.216) 0.062 (0.102) 0.865 (0.302) 0.271 (0.161) Alternative Education Placement 0.020 (0.004) 0.005 (0.002) 1,740,282 Value-Added Attendance (4) 1.132 (0.544) 2.338 (0.757) 0.047 (0.167) 1.852 (0.364) 0.199 (0.174) 1.873 (0.367) 0.181 (0.178) 3.071 (1.129) 0.032 (0.171) 3.242 (1.203) 0.061 (0.172) 4.721 (1.646) 0.143 (0.212)
427
Number of Observations 1,777,994 1,777,994
Drop year prior to charter Drop 2 years prior to charter
Kyriazidou (1997) attrition model
First differences (unweighted) MSE minimizing bandwidth 1/4 MSE minimizing bandwidth
A: Disciplinary Infractions and Attendance 1.029 2,233,050 0.356 (0.127) (0.470) 2.268 2,233,050 Start-up 0.795 (0.113) (0.607) Conversion 0.023 0.044 (0.070) (0.186) 2.149 2,221,517 Start-up 0.789 (0.101) (0.545) Conversion 0.028 0.045 (0.074) (0.182) 2.058 2,211,570 Start-up 0.775 (0.088) (0.491) Conversion 0.026 0.025 (0.078) (0.177) 2.467 1,457,716 Start-up 0.809 (0.171) (1.027) Conversion 0.029 0.017 (0.073) (0.194) 2.500 1,457,716 Start-up 0.818 (0.174) (1.015) Conversion 0.034 0.017 (0.075) (0.194) 3.125 1,457,716 Start-up 0.879 (0.190) (1.218) Conversion 0.109 0.013 (0.095) (0.218) B: Additional Outcomes (Levels Only) Any Infraction 0.250 (0.036) 0.002 (0.015) 2,233,050 Expulsion 0.0036 (0.0013) 0.0016 (0.0007) 1,740,282
1,758,480
1,756,982
1,067,566
1,067,566
1,067,566
Start-up Convert Observations
Retention 0.0024 (0.0014) 0.0001 (0.0001) 1,777,994
All regressions include an individual xed effect, free lunch status, reduced-price lunch status, other economic disadvantage, recent immigration status, whether a parent is a migrant worker, and grade-by-year indicators. Standard errors clustered by school in parentheses. , , denote signicance at the 10%, 5%, and 1% levels, respectively. Levels models cover 19942006 and grades 112 and value-added models 19942006 and grades 212 for discipline and attendance. Because students are dened to be in the sample in the rst grade and year of each sample, these observations are dropped from the Kyriazidou attrition analysis. Additional outcomes cover all grades 112 but retention, expulsion, and alternative education program placement are not available in all years.
ghting by 0.024 (s.e. 0.009).30 It is unlikely that principals punish students for these infractions with punishments that are less severe than in-school suspensions; hence, improvements in these outcomes are most likely due to behavioral improvements.31 Third, at four to seven times the standard error, the enforcement or reporting bias would need to be very large to make the discipline estimates statistically insignicant. Fourth, a multinomial logit regression of type of punishment on charter status shows that start-up charter students were more likely to receive out-of-school suspensions than in-school suspensions, suggesting that if anything, punishments in start-up charters are harsher than in noncharters. It is possible that
standard errors are clustered by school. Value-added models were not statistically signicant. However, since these are low-frequency events, it is very rare to get more than one infraction in subsequent years; thus, it would be difcult for value-added models to identify impacts on these outcomes. 31 In addition, for substance abuse and criminal infractions, the principal is legally obligated to notify the police department.
30 All
this result reects a drop in minor infractions without the commensurate drop in major infractions. However, the drop in severe infractions suggests this is unlikely to be the case.32 Finally, in panel B of table 3, the rst three columns provide results for some alternative measures of discipline. Since these are all binary outcomes, I report only levels models. All regressions include the same covariates as in the baseline regressions. I consider the effect of attending a charter on whether the student has any disciplinary infractions during the year, whether the student is expelled, and whether the student is placed in an alternative school for students with disciplinary problems. Start-up charters provide statistically signicant drops in likelihood of all three of these outcomes.
32 The unit of observation is at the student infraction level, and the leftout category is in-school suspension. The regression includes covariates for gender, race, free lunch, reduced-price lunch, other economic disadvantage, immigration status, whether student is a recent immigrant, whether the parent is a migrant worker, infraction typesubstance abuse, violent crime, nonviolent crime, truancy, otherand grade-by-year dummies. The estimate of attending a start-up charter on out-of-school suspensions is statistically signicant at the 5% level. These results are available on request.
428

Table 4.Charter Impacts by Grade Level and Age of Charter Math (1) Levels Startup Grades 15 Startup Grades 68 Startup Grades 912 Observations Value added Startup Grades 15 Startup Grades 68 Startup Grades 912 Observations Startup 1 Year Startup 2 Years Startup 3 Years Startup 4 Years Startup 5+ Years Convert 1 Year Convert 2 Years Convert 3 Years Convert 4 Years Convert 5+ Years Observations Reading (2) Language (3) Discipline (4) Attendance (5)
A: Impacts by Grade Level 0.010 (0.042) 0.182 (0.064) 0.021 (0.095) 1,141,480 0.011 (0.056) 0.071 (0.023) 0.024 (0.049) 779,343 0.046 (0.045) 0.073 (0.059) 0.024 (0.065) 0.185 (0.081) 0.056 (0.060) 0.093 (0.071) 0.086 (0.055) 0.149 (0.059) 0.123 (0.099) 0.033 (0.054) 1,141,480 0.005 (0.033) 0.064 (0.028) 0.040 (0.096) 1,141,480 0.011 (0.035) 0.076 (0.022) 0.029 (0.077) 1,141,480 0.182 (0.035) 0.933 (0.122) 0.916 (0.156) 2,233,050 0.333 (0.102) 0.677 (0.238) 0.401 (0.198) 1,777,994 0.653 (0.098) 0.621 (0.123) 0.691 (0.122) 0.858 (0.150) 0.921 (0.127) 0.033 (0.062) 0.097 (0.042) 0.070 (0.113) 0.015 (0.061) 0.093 (0.119) 2,232,727 0.029 (0.103) 1.607 (0.262) 3.562 (0.623) 2,233,050 0.231 (0.224) 1.229 (0.302) 3.734 (0.868) 1,777,994 5.898 (2.150) 1.820 (1.829) 1.646 (0.438) 3.320 (1.020) 1.148 (0.735) 0.025 (0.209) 0.114 (0.171) 0.211 (0.274) 0.063 (0.246) 0.170 (0.238) 2,232,727
0.058 0.045 (0.036) (0.038) 0.005 0.010 (0.023) (0.021) 0.076 0.014 (0.053) (0.034) 779,343 779,343 B: Impacts by Age of Charter (Levels) 0.040 0.031 (0.044) (0.027) 0.013 0.057 (0.060) (0.057) 0.023 0.078 (0.045) (0.030) 0.045 0.061 (0.037) (0.034) 0.031 0.012 (0.050) (0.044) 0.038 0.017 (0.034) (0.043) 0.055 0.034 (0.041) (0.040) 0.077 0.117 (0.038) (0.057) 0.086 0.082 (0.094) (0.056) 0.051 0.005 (0.049) (0.036) 1,141,480 1,141,480
Each column in each panel is a separate regression. All regressions include an individual xed effect, free lunch status, reduced-price lunch status, other economic disadvantage, recent immigration status, whether a parent is a migrant worker, and grade-by-year indicators. Grade levels refer to the grade of the student. Standard errors clustered by school in parentheses. , , denote signicance at the 10%, 5%, and 1% levels, respectively. Levels models cover 19982006 and grades 111 for test scores and 19942006 and grades 112 for other outcomes. Value-added models cover 19992006 and grades 211 for test scores and 19942006 and grades 212 for other outcomes. Columns in each panel show separate regressions. Grade-level regressions also contain an indicator for conversion charters, but since only one conversion covers grades 78 and no conversions cover grades 912, I do not separate that estimate by grade level. Value-added results for age of charter are similar to levels results and are available in the online appendix. Observation counts by school type are also available in the online appendix.
Conversion charters also show drops in expulsion rates and alternative education placement, but these are substantially smaller than for start-ups. Thus, the discipline results for startups are robust across multiple margins. The last column of panel B looks at retention. The results show a slight increase in retentions for start-ups, but this is signicant only at the 10% level. Whether students are retained more in charters is potentially interesting but provides an unclear interpretation. Higher retention could indicate that students perform worse, but it could also indicate a policy difference where charters are more likely to hold marginal students back or that charters are better at identifying students who need to be held back. So far I have established that on average, start-up charters provide little improvement in cognitive skills while generating large improvements in noncognitive skills. Since this is an average result, these impacts may vary by school and student characteristics. Table 4 provides some results on how
charter impacts differ by grade level and age of the charter. The table shows total effects for each type of school. Models with main effects and interactions are available in the online appendix. Panel A looks at variation by grade level. Since the there is only one conversion charter in the sample with grades 7 and 8 and none with grades 9 through 12, a pooled estimate for conversion charters is included in the regression but not shown.33 In columns 1 through 3, I show that while elementary and high school start-ups have no statistically signicant impact on any test score, start-ups with grades 6 through 8 (middle school grades) fare quite well. All three of the estimates in levels models show that students in these grades perform statistically signicantly better than noncharter students, and models with main effects and interactions
33 Full results, along with observation counts by school type, are provided in the online appendix.

Table 5.Start-up Charter Impacts by Student Characteristics Math (1) Start-up Start-up Black Start-up Hispanic Start-up Female Start-up Economically Disadvantaged Start-up Immigrant Observations Start-up Start-up Black Start-up Hispanic Start-up Female Start-up Economically Disadvantaged Start-up Immigrant Observations 0.286 (0.185) 0.238 (0.166) 0.305 (0.154) 0.056 (0.027) 0.028 (0.025) 0.034 (0.030) 1,141,480 0.027 (0.061) 0.054 (0.061) 0.024 (0.038) 0.034 (0.019) 0.041 (0.029) 0.053 (0.027) 779,343 Reading (2) A: Levels 0.086 (0.086) 0.087 (0.076) 0.133 (0.045) 0.005 (0.022) 0.024 (0.013) 0.030 (0.019) 1,141,480 B: Value added 0.102 (0.058) 0.093 (0.060) 0.011 (0.056) 0.034 (0.019) 0.014 (0.040) 0.034 (0.027) 779,343 Language (3) 0.024 (0.074) 0.034 (0.058) 0.006 (0.027) 0.013 (0.019) 0.020 (0.017) 0.018 (0.025) 1,141,480 0.158 (0.048) 0.196 (0.044) 0.141 (0.044) 0.042 (0.024) 0.003 (0.022) 0.033 (0.026) 779,343 Discipline (4) 0.948 (0.120) 0.228 (0.119) 0.098 (0.093) 0.238 (0.074) 0.039 (0.043) 0.078 (0.033) 2,233,050 0.745 (0.219) 0.075 (0.161) 0.052 (0.072) 0.381 (0.151) 0.071 (0.093) 0.058 (0.042) 1,777,994 Attendance (5) 2.601 (0.956) 1.094 (0.740) 0.434 (0.467) 0.055 (0.244) 0.090 (0.268) 1.358 (0.565) 2,233,050 3.381 (1.404) 1.908 (1.506) 0.060 (1.150) 0.329 (0.429) 0.021 (0.540) 0.409 (0.163) 1,777,994
429
Each column in each panel is a separate regression. All regressions include an individual xed effect, free lunch status, reduced-price lunch status, other economic disadvantage, recent immigration status, whether a parent is a migrant worker, and grade-by-year indicators. Regressions also include the main effect and the same set of interactions for conversion charters, along with start-up and conversion interactions with other nonwhite students. These students make up less than 1% of start-up and 4% of conversion charter students. Full results are available in the online appendix. Standard errors clustered by school in parentheses. , , denote signicance at the 10%, 5%, and 1% levels, respectively. Levels models cover 19982006 and grades 111 for test scores and 19942006 and grades 112 for other outcomes. Value-added models cover 19992006 and grade 211 for test scores and 19942006 and grades 212 for other outcomes.
show they perform better than both their elementary and highschool counterparts. However, for value-added models, only math has a statistically signicant effect. These results imply that middle school start-ups increase math scores by 0.7 to 0.18 standard deviations. Reading and language increase by at most 0.07 and 0.08 standard deviations, but I cannot rule out a zero effect. Hence, while on average there is little evidence of cognitive skill improvements from start-up charters, there is some weak evidence of improvements for middle schools. Discipline and attendance results show evidence of noncognitive skill improvements at all grade levels. Elementary schools show improvement in discipline but not attendance, while other grade levels show improvements in both measures. Nonetheless, average attendance rates are higher in elementary grades so there is less room for improvement. The district-wide average is 96.4% for grades 1 through 5 and 92.6% for grades 6 and higher. Panel B of table 4 looks at how charter impacts vary by the age of the charter in levels models. Value-added models provided similar results and are available in the online appendix. Previous work on charter schools has generally found that as charters age, their test score impacts improve (Booker et al., 2007; Hanushek et al., 2007; Sass, 2006; Bifulco & Ladd, 2006). While I caution that since I have 31 charters, there is less variation in this school-level characteristic than in previous papers, I nonetheless nd little evidence of improvement in test scores for start-ups or conversions as the schools
age. I do nd some evidence of improvement in discipline in start-ups, as those that are ve or more years old have 0.3 fewer annual disciplinary infractions per student than rstyear start-ups. However, this is offset by lower attendance improvements after the rst year. It is unclear why attendance rates would drop after the rst year. One possibility is that schools that are new need to maintain a closely knit community in order to succeed, so they try harder to induce students to attend. In table 5, I consider how charter school impacts vary by student characteristics. Since I do not nd substantial evidence of impacts for conversion charters, I report only the estimates for start-up charters.34 Each column in panel A or B is a separate regression with a main effect and interaction effects. No estimate for race is signicantly different from the main effect in both levels and value-added models, so I cannot say for sure if there is a racial difference. There are, however some notable differences by gender. Girls in start-ups appear to perform better on math tests than boys by 0.03 to 0.06 standard deviations. While somewhat tenuous the value-added estimate is signicant only at 10%this is an intriguing result as there has been considerable focus in education policy on how to improve math scores for girls. Disciplinary infraction impacts are smaller for girls, though they
34 Full results, including conversion estimates, are provided in the online appendix.
430

Table 6.Mechanisms of Discipline and Attendance Impacts in Start-up Charters A: Levels Disciplinary Infractions Attendance B: Value-Added Disciplinary Infractions Attendance 0.498 (0.151) 1,777,994 0.536 (0.139) 1,769,243 0.586 (0.183) 1,770,359 0.520 (0.205) 1,769,243 0.354 (0.144) 1,769,243 0.341 (0.184) 1,777,666 0.125 (0.200) 1,777,666 2.338 (0.757) 1,777,994 2.406 (0.791) 1,769,243 2.276 (0.759) 1,770,359 2.073 (0.756) 1,769,243 1.823 (0.733) 1,769,243 2.099 (0.680) 1,777,666 2.277 (0.837) 1,777,666
(1) (2) (3) (4) (5) (6) (7)
Baseline (from tables 2 and 3) Controlling for per student instructional expenditures Controlling per student school leadership expenditures Controlling per student other expenditures Controlling for per student instructional, school leadership, and other expenditures Controlling for student body composition Controlling for enrollment
Coefcient Observations Coefcient Observations Coefcient Observations Coefcient Observations Coefcient Observations Coefcient Observations Coefcient Observations
0.795 (0.113) 2,233,050 0.799 (0.121) 2,221,995 0.852 (0.131) 2,223,425 0.849 (0.129) 2,221,995 0.755 (0.115) 2,221,995 0.798 (0.113) 2,232,700 0.550 (0.129) 2,232,700
2.268 (0.607) 2,233,050 2.318 (0.582) 2,221,995 1.908 (0.583) 2,223,425 1.716 (0.613) 2,221,995 1.733 (0.641) 2,221,995 2.343 (0.582) 2,232,700 2.899 (0.706) 2,232,700
Each coefcient is from a separate regression. All regressions include an individual xed effect, free lunch status, reduced-price lunch status, other economic disadvantage, recent immigration status, whether a parent is a migrant worker, and grade-by-year indicators. Regressions also include a conversion charter estimate that is not shown here. Additional controls for potential charter mechanisms are added as described above. Each row/column combination is a separate regression with both conversion and start-up charter estimates. Full results are provided in the online appendix. Standard errors clustered by school in parentheses. , , denote signicance at the 10%, 5%, and 1% levels, respectively.
still show large improvements. This is not surprising, as boys tend to have more discipline problems on average than girls do. For economic status, there is no statistically signicant difference for any outcome. Immigrants have lower improvements in attendance than nonimmigrants, but the combined main effect and interaction is still positive.35 In table 6 I investigate some mechanisms the charter impacts may work through. I showed in table 1 that startup charters have more instructional expenditures and fewer expenditures on other functions, and they are generally smaller than noncharter schools. Hence, seeing how the impact estimates change as we control for these characteristics can provide insight into the paths through which charters affect students. Nonetheless, these covariates are potentially endogenous, so one should be careful of interpreting any change in the charter impact estimate as being causally determined by the added covariates. Tests score estimates for all charters and discipline and attendance estimates for conversions are generally unaffected by adding these covariates, so I leave those results in the online appendix. The rst row of table 6 repeats the baseline estimates from tables 2 and 3. The second through fth rows show what happens when we add different categories of per student
expenditures. Adding these factors does not generate substantial changes in the impact estimates. In row 6 I control for the percentage of each school that is white, Hispanic, black, limited English procient, special education, gifted, and economically disadvantaged. In this case, there is a drop in the discipline estimate in value-added models, to the point where it is signicant only at the 10% level. However, all of the other estimates change little. Row 7 controls for a quadratic in total school enrollment. Since start-ups tend to be smaller than noncharters, administrators may nd it easier to maintain control over students and spend more time dealing with discipline problems. The results in row 7 are consistent with this theory. When enrollment is added, the disciplinary infraction estimate rises by 0.145 in levels and .373 in value-added models. While the levels estimate remains statistically signicant, the valueadded estimates becomes insignicant. Therefore, it appears that school size and peer characteristics may play roles in the discipline improvements found in the start-up charters. In particular, part of the discipline improvement may be due to administrators having smaller schools that they can closely monitor. This could be an important avenue for further study as there is much interest in the benets of small schools.
B. Evolution of Charter Impacts over Time
additional element of heterogeneity that one may expect to see is that students with precharter behavioral problems improve more in startup charters than other students do. Indeed, regressions that interact charter status with infractions in the year prior to charter entry show that the impacts on infractions increase as precharter infractions increase. Nonetheless, this does not appear to spill over into test scores, as improvements in these measures fall with the number of infractions, though not to the point where a large number of students would have signicantly negative test score impacts. These results are available on request.
35 An
One of the key advantages of the LUSD-SW data set is that the length and breadth of the panel allows me to investigate some long-term impacts of charters on cognitive and noncognitive skills that most of the previous research is not able to do. First, in table 7, I provide results from regressions where charter impacts are allowed to vary by time spent in

Table 7.Charter Impacts by Time in Charter Math (1) Start-up1 year Start-up2 years Start-up3 years Start-up4 or more years Conversion1 year Conversion2 years Conversion3 years Conversion4 or more years Observations Start-up1 year Start-up2 years Start-up3 years Start-up4 or more years Conversion1 year Conversion2 years Conversion3 years Conversion4 or more years Observations 0.013 (0.060) 0.008 (0.078) 0.114 (0.098) 0.151 (0.270) 0.095 (0.050) 0.013 (0.138) 0.062 (0.139) 0.019 (0.137) 1,141,480 0.053 (0.052) 0.005 (0.068) 0.159 (0.163) 0.446 (0.183) 0.069 (0.108) 0.163 (0.155) 0.335 (0.165) 0.012 (0.332) 779,343 Reading (2) A: Levels 0.016 (0.045) 0.062 (0.075) 0.128 (0.106) 0.122 (0.150) 0.056 (0.029) 0.074 (0.100) 0.031 (0.093) 0.000 (0.142) 1,141,480 B: Value added 0.006 (0.039) 0.084 (0.050) 0.103 (0.112) 0.202 (0.120) 0.021 (0.075) 0.179 (0.112) 0.165 (0.139) 0.001 (0.218) 779,343 Language (3) 0.004 (0.042) 0.008 (0.059) 0.122 (0.082) 0.042 (0.130) 0.064 (0.039) 0.007 (0.091) 0.020 (0.061) 0.095 (0.086) 1,141,480 0.046 (0.047) 0.017 (0.048) 0.121 (0.089) 0.375 (0.148) 0.053 (0.055) 0.018 (0.087) 0.253 (0.223) 0.255 (0.171) 779,343 Discipline (4) 0.837 (0.115) 1.042 (0.200) 0.981 (0.168) 1.577 (0.354) 0.006 (0.048) 0.075 (0.082) 0.182 (0.111) 0.207 (0.118) 2,233,050 0.792 (0.222) 0.226 (0.122) 0.232 (0.359) 0.035 (0.731) 0.004 (0.063) 0.186 (0.093) 0.208 (0.097) 0.150 (0.094) 1,777,994 Attendance (5) 3.270 (1.009) 2.197 (0.412) 3.756 (0.660) 11.774 (10.629) 0.029 (0.179) 0.175 (0.360) 0.564 (0.421) 0.984 (0.535) 2,233,050 3.800 (1.408) 1.700 (0.442) 3.645 (0.636) 7.327 (5.546) 0.148 (0.203) 0.477 (0.329) 0.740 (0.501) 0.992 (0.604) 1,777,994
431
Each column in each panel is a separate regression. All regressions include an individual xed effect, free lunch status, reduced-price lunch status, other economic disadvantage, recent immigration status, whether a parent is a migrant worker, and grade-by-year indicators. All post-rst-year periods are instrumented by the potential number of years the student could have been in the charter school based on grade at charter entry. Standard errors clustered by school in parentheses. , , denote signicance at the 10%, 5%, and 1% levels, respectively. Levels models cover 19982006 and grades 111 for test scores and 19942006 and grades 112 for other outcomes. Value-added models cover 19992006 and grades 211 for test scores and 19942006 and grades 212 for other outcomes. Observation counts by time in charter are provided in the online appendix.
a charter. It is possible that initially a student may not perform well in a charter due to the shock of having to change educational environments, but over time, the student could acclimate to the new environment and improve. Or it is possible that the new environment of attending a charter could temporarily motivate the student and improve performance, but as time goes on, the student may return to old habits. In order to address the potential for endogenous exit, I instrument for charter attendance in any year after the rst. In this case, I use the number of potential years a student could have been in the charter. That is, I subtract the rst grade offered in the charter from the grade the student is predicted to be in. The predicted grade is calculated by assuming normal grade progression from the grade the student was in when he or she entered the charter. For test scores (columns 13), start-up charters show some evidence of improvement when students have been in a charter for four or more years, but only in value-added models. Thus, there appears to be some weak evidence that students who spend long periods of time in start-ups improve, but those who spend only short periods of time in start-ups do no better or worse. For
conversion charters, it appears that test score impacts remain roughly constant as time in the charter increases.36 Looking at columns 4 and 5, for discipline and attendance there appears to be improvement as time in start-up charters increase. For example, in the rst year, annual disciplinary infractions are down by 0.8 in levels models. After the third year, they are down by 1.6. An F-test of this difference is statistically signicant at the 5% level. Attendance also shows a pattern of improvement. The pattern for discipline does not hold for start-ups in value-added models. This suggests that discipline improves substantially in the rst year, but further improvements in subsequent years are small. Nonetheless, discipline impacts do increase as time spent in the charter increases. These results also show that there is no mean reversion in charter impacts and thus provide further evidence that the preentry dips in test scores and attendance do not impose substantial bias.
36 The 2SLS models are somewhat different from OLS models, which show improvement in start-up charters over time. These results are available in the online appendix.
432

Figure 6.Test Scores before and after Charter Exit
Figure 7.Discipline and Attendance by before and after Charter Exit

Table 8.Persistence of Charter Impacts Math (1) Start-up 1 to 3 years after start-up 4 or more years after start-up Conversion 1 to 3 years after conversion 4 or more years after conversion F-test of post-start-up estimates F-test of post-conversion estimates Observations Start-up 1 to 3 years after start-up 4 or more years after start-up Conversion 1 to 3 Years after conversion 4 or more years after conversion F-test of post-start-up estimates F-test of post-conversion estimates Observations 0.038 (0.046) 0.048 (0.041) 0.040 (0.165) 0.007 (0.039) 0.050 (0.058) 0.065 (0.063) 2.11 5.69 1,141,480 0.002 (0.033) 0.033 (0.041) 0.073 (0.187) 0.072 (0.063) 0.109 (0.072) 0.132 (0.070) 0.52 0.74 779,343 Reading (2) A: Levels Models 0.734 (0.109) 0.167 (0.047) 0.161 (0.051) 0.015 (0.066) 0.025 (0.052) 0.027 (0.052) 0.10 2.65 1,141,480 B: Value added Models 0.446 (0.154) 0.192 (0.052) 0.025 (0.046) 0.092 (0.058) 0.185 (0.051) 0.056 (0.041) 3.29 2.04 779,343 Language (3) 0.734 (0.123) 0.244 (0.151) 0.104 (0.165) 0.110 (0.120) 0.173 (0.198) 0.053 (0.166) 0.96 0.57 1,141,480 0.396 (0.148) 0.561 (0.154) 0.237 (0.193) 0.092 (0.115) 0.155 (0.173) 0.030 (0.119) 0.51 2.18 779,343 Discipline (4) 2.027 (0.554) 0.712 (0.209) 0.829 (0.224) 0.322 (0.189) 0.522 (0.191) 0.358 (0.209) 2.93 0.96 2,233,050 2.319 (0.776) 0.001 (0.152) 0.383 (0.181) 0.062 (0.170) 0.144 (0.150) 0.316 (0.170) 13.04 0.79 1,777,994 Attendance (5) 2.509 (0.725) 0.655 (0.521) 0.033 (1.253) 0.771 (0.289) 1.295 (0.429) 1.302 (0.432) 1.3 6.42 2,233,050 2.414 (0.853) 0.027 (0.438) 0.602 (0.910) 0.708 (0.231) 1.255 (0.278) 1.099 (0.311) 0.30 11.59 1,777,994
433
Each column in each panel is a separate regression. Two- and three-year postcharter estimates are similar to the one- and four-year estimates, so I do not show them here, though the full set of results is provided in the online appendix. All regressions include an individual xed effect, free lunch status, reduced-price lunch status, other economic disadvantage, recent immigration status, whether a parent is a migrant worker, and grade-by-year indicators. All post-charter periods are instrumented by whether the student is the listed number of years beyond the last grade level covered by his or her prior charter school. Standard errors clustered by school in parentheses. , , denote signicance at the 10%, 5%, and 1% levels, respectively. Levels models cover 19982006 and grades 111 for test scores and 19942006 and grades 112 for other outcomes. Value-added models cover 19992006 and grades 211 for test scores and 19942006 and grades 212 for other outcomes. Observation counts by post-charter period are provided in the online appendix.
I now turn to how charter impacts persist after students return to regular public schools. Analyzing persistence informs us as to whether the skills students learn in charters remain without the need for reinforcement. Persistence can also generate bias in the impact estimates since some of the impacts would be attributed to noncharter periods. Figures 6 and 7 provide an initial look at whether charter impacts persist by graphing residuals from the regressions used for gures 4 and 5 for students before and after they leave charters. Figure 6 shows that after charter exit, math test scores drop for start-ups and all subjects drop for conversions. Figure 7 shows that the discipline and attendance gains achieved while students are enrolled in start-ups drop off after they leave. These measures are relatively stable after students leave conversions. Hence, these gures suggest that charter impacts do not persist after students leave. In table 8, I provide regression results that account for the persistence of charter impacts when students return to regular public schools. The table shows 2SLS results using the instrument for post-charter periods described at the end of section IIIB. To check persistence, I use indicators for having attended a charter one to three years prior and for
having attended a charter more than four years prior.37 Test scores are provided in columns 1 to 3, discipline in column 4, and attendance in column 5. Each column in panels A and B refers to a single regression. First, for conversion charters, any signicant positive impacts found in the baseline models become statistically insignicant. This provides additional conrmation that conversion charters are ineffective at improving student outcomes. Turning to the results for start-ups, columns 1, 2 and 3 show that after leaving the charter, students perform well on test scores in the rst three years, with statistically signicant and positive impacts in the levels models for math. Nonetheless, all of the other estimates for post-start-up test scores are either negative or statistically insignicant. Thus, the evidence for persistence in test scores for start-ups is weak. As such, the estimates for the start-up impact on test scores remain statistically insignicant at the 5% or lower level in all models.
37 OLS results are available in the online appendix. Analyses using indicators for one, two, three, and four or more years post-charter showed similar results and are also available in the online appendix.
434
Columns 4 and 5 show that after start-up students return to regular public schools, disciplinary infractions increase and attendance falls to the point where they do not differ from noncharter students. While it is true that in the post-charter periods, the students are older and thus we would expect these outcomes to worsen, the inclusion of grade-by-year indicators corrects for this problem. Hence, there is no persistent impact of start-up charters on noncognitive skills. One possible explanation for this result is that the noncognitive skill improvements from start-up charters need continual reinforcement or else they dissipate. Another possible explanation is that the improvements in behavior are due to institutional factors rather than skill formation. Hence, while the start-ups provide an environment that helps students improve behavior and attendance, once they return to regular public schools, they lose the benet of that environment. A third alternative is that the estimated impacts reect reporting and enforcement bias in charters. However, as discussed in section IVA, it is highly unlikely that attendance rates are manipulated while the discipline results appear to be, at least partially, due to real behavioral changes.
V. Conclusion
In this paper, I investigate the effect of charter schools on students who attend them using data from a large urban school district in the Southwest (LUSD-SW) with an extensive charter program. I provide a broad outlook on charter schools by focusing on multiple outcomes and using a panel with wide grade coverage and long time frames to provide analyses of long-term impacts both while in and after attending charters. By combining analyses of test scores with discipline and attendance impacts, I am able to identify whether charter schools affect both cognitive and noncognitive skill formation. I nd that while charters that convert from regular public schools to charter status have little effect on cognitive or noncognitive skills, schools that begin as charters, that is, start-up charters, generate large improvements in discipline and attendance. Disciplinary infractions drop by 0.5 to 0.8 instances per year compared to a precharter mean of 1.1, and attendance rates rise by 2.3 percentage points, which is equivalent to 26% of the precharter mean. These impacts are statistically signicant at the 1% level across many models and specication checks. While discipline and attendance impacts could also be caused by cognitive skill enhancements, I can establish improvements in test scores only for math in middle school start-ups, whereas estimates for elementary and high schools are close to zero on all test score measures. Hence, I interpret the discipline and attendance results as showing that start-up charters generate noncognitive skill improvements. These individual xed-effects results are robust to potential biases from selection into charters off trends in outcomes and attrition. The long and broad nature of my panel also allow me to follow students for a number of years while they are
enrolled in charters and after they leave charters. While impacts for either type of charter are relatively stable as time in the charter increases, the discipline and attendance impacts found in start-ups do not persist after students return to regular public schools, so there do not appear to be longterm behavioral improvements. This result suggests that any noncognitive skills that have improved in start-up charters require constant reinforcement. I cannot rule out, however, that the improvements in attendance and discipline may instead reect institutional differences between charters and noncharter public schools. I also caution that while the discipline and attendance results are intriguing, they may reect reporting or enforcement differences. The lack of persistence increases this concern as such a result is consistent with reporting bias. Since attendance is difcult to manipulate or misreport, this is mainly a concern for discipline. To address this, I provide multiple pieces of indirect evidence suggesting that the discipline results, at least in part, reect real behavioral changes in the start-up charter students. In particular, the large impacts on attendance buttress the discipline improvements, and combined, they provide substantial evidence that startups generate noncognitive skill improvements in spite of the lack of cognitive skill improvement. Finally, I should note that charter schools vary considerably in their mission and the rules they are subject to; each is a unique entity. Hence, it is possible that the results for the charters in LUSD do not extend to other school districts and states. Therefore, it is important that future research on charter schools consider alternative outcomes like discipline and attendance so that we can determine whether noncognitive skill improvement is a general characteristic of charter schools.
REFERENCES Ashenfelter, Orley, Estimating the Effect of Training Programs on Earnings, this review 60:1 (1978), 4757. Betts, Julian R., Lorien A. Rice, Andrew C. Zau, Y. Emily Tang, and Cory R. Koedel, Does School Choice Work? Effects on Student Integration and Achievement, Public Policy Institute of California report (2006). Bifulco, Robert, and Helen F. Ladd, The Impacts of Charter Schools on Student Achievement: Evidence from North Carolina, Education Finance and Policy 1:1 (2006), 123138. Booker, Kevin, Scott M. Gilpatric, Timothy Gronberg, and Dennis Jansen, The Impact of Charter School Attendance on Student Performance, Journal of Public Economics 91:5/6 (2007), 849876. Buddin, Richard, and Ron Zimmer, Student Achievement in Charter Schools: A Complex Picture, Journal of Policy Analysis and Management 24:2 (2005), 351371. Chay, Kenneth Y., Patrick J. McEwan, and Miguel Urquiola, The Central Role of Noise in Evaluating Interventions That Use Test Scores to Rank Schools, American Economic Review 95:4 (2005), 12371258. Clark, Damon, The Performance and Competitive Effects of School Autonomy, Journal of Political Economy 117:4 (2009), 745783. Dustmann, Christian, and Maria E. Rochina-Barrachina, Selection Correction in Panel Data Models: An Application to Labour Supply and Wages, IZA discussion paper no. 162 (2000). Guryan, John, Desegregation and Black Dropout Rates, NBER working paper no. 8345 (2001).

Hanushek, Eric A., John F. Kain, Steven G. Rivkin, and Gregory F. Branch, Charter School Quality and Parental Decision Making with School Choice, Journal of Public Economics 91:5/6 (2007), 823848. Heckman, James J., and Jeffrey A. Smith, The Pre-Programme Earnings Dip and the Determinants of Participation in a Social Programme: Implications for Simple Programme Evaluation Strategies, Economic Journal 109:457 (1999), 313348. Heckman, James J., and Yona Rubinstein, The Importance of Noncognitive Skills: Lessons from the GED Testing Program, American Economic Review: Papers and Proceedings 91:2 (2001), 145149. Heckman, James J., Jora Stixrud, and Sergio Urzua, The Effects of Cognitive and Noncognitive Abilities on Labor Market Outcomes and Social Behavior, Journal of Labor Economics 24:3 (2006), 411482. Hoxby, Caroline M., and Sonali Murarka, Charter Schools in New York City: Who Enrolls and How They Affect Their Students Achievement, NBER working paper no. 14852 (2009). Hoxby, Caroline M., and Jonah E. Rockoff, The Impact of Charter Schools on Student Achievement, Department of Economics, Harvard University, mimeo. (2004). Kyriazidou, Ekaterini, Estimation of a Panel Data Sample Selection Model, Econometrica 65:6 (1997), 13351364. McClure, Larry, Betsy Strick, Rachel Jacob-Almeida, and Christopher Reicher, The Preuss School at UCSD: School Characteristics and Students Achievement, CREATE-UCSD report (2005). Sass, Tim R., Charter Schools and Student Achievement in Florida, Education Finance and Policy 1:1 (2006), 123138. Segal, Carmit, Classroom Behavior, Journal of Human Resources 43:4 (2009a), 783814. Misbehavior, Education, and Labor Market Outcomes, Department of Economics, Universitat Pompeu Fabra, mimeo. (2009b). Solmon, Lewis, and Pete Goldschmidt, Comparison of Traditional Public Schools and Charter Schools on Retention, School Switching, and Achievement Growth, Goldwater Institute policy report (2004). Solmon, Lewis, Kern Paark, and David Garcia, Does Charter School Attendance Improve Test Scores? The Arizona Results, Goldwater Institute occasional report (2001). Todd, Petra E., and Kenneth I. Wolpin, The Production of Cognitive Achievement in Children: Home, School and Racial Test Score Gaps, Journal of Human Capital 1:1 (2007), 91136. Weiher, Gregory, R., and Kent L. Tedin, Does Choice Lead to Racially Distinctive Schools? Charter Schools and Household Preferences, Journal of Policy Analysis and Management 21 (2000), 7992. Zimmer, Ron, and Richard Buddin, Academic Outcomes (pp. 3762), in Charter School Operations and Performance: Evidence from California (Santa Monica, CA: RAND, 2003).
435
Charter School Performance in Two Large Urban Districts, Journal of Urban Economics 60:2 (2006), 307326. Zimmer, Ron, Suzanne Blanc, Brian Gill, and Jolley Christman, Evaluating the Performance of Philadelphias Charter Schools, Rand working paper (2008).
APPENDIX
Proof of Expected Value of Level and Value-Added Fixed Effects Estimates Bounding the Lagged-Dependent Variable Model with Fixed Effects Let us rst simplify notation and denote X as a k nt vector of demeaned covariates while Y is a 1 nt vector of the demeaned student outcome variable and Yt1 is the 1 nt vector of demeaned once-lagged outcome variables. Our true model becomes Yt = X + Yt1 + . (A1)
In a levels framework, the lagged outcomes enter into the error term such that there is composite error: = Yt1 + . This provides us with E(L ) = + [X X]1 [X E(Yt1 )]. (A3) (A2)
For a value-added model, we subtract Yt1 from each side of equation (A1) to get Yt Yt1 = X + ( 1)Yt1 + , which will provide us with an estimate of b such that E(VA ) = + ( 1)[X X]1 [X E(Yt1 )]. (A5) (A4)
Let us further dene the matrix A = [X X]1 [X E(Yt1 )] and the kth row of A as Ak ; hence, E(L ) = k + Ak k E(VA ) = k + ( 1)Ak .
k
(A6) (A7)
Thus, assuming that 0 1, if Ak > 0, then E(L ) > > E(VA ), k k while if Ak < 0, then E(L ) < < E(VA ). In either case, the levels model k k and value-added models bound .
This article has been cited by: 1. C. Kirabo Jackson. 2012. School competition and teacher labor markets: Evidence from charter school entry in North Carolina. Journal of Public Economics 96:5-6, 431-448. [CrossRef] 2. Elisabetta Gentile, Scott A. Imberman. 2011. Dressed for Success? The Effect of School Uniforms on Student Achievement and Behavior. Journal of Urban Economics . [CrossRef] 3. Celeste K. Carruthers. 2011. New schools, new students, new teachers: Evaluating the effectiveness of charter schools. Economics of Education Review . [CrossRef] 4. Ron Zimmer, Brian Gill, Kevin Booker, Stphane Lavertu, John Witte. 2011. Examining charter student achievement effects across seven states. Economics of Education Review . [CrossRef]

Imberman 2011 - Achievement and Behavior in Charter Schools

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Imberman 2011 - Achievement and Behavior in Charter Schools

Uploaded by

Copyright:

Available Formats

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS: DRAWING A MORE COMPLETE PICTURE

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

THE REVIEW OF ECONOMICS AND STATISTICS

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

THE REVIEW OF ECONOMICS AND STATISTICS

Noncharter 0.49 (0.50)

cognitive skills. To conrm this result, I turn to regression analysis.

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

THE REVIEW OF ECONOMICS AND STATISTICS

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

yit = + S,0 Startupit +

A. Charter School Impacts on Student Outcomes

THE REVIEW OF ECONOMICS AND STATISTICS

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

THE REVIEW OF ECONOMICS AND STATISTICS

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

Number of Observations 1,777,994 1,777,994

Drop year prior to charter Drop 2 years prior to charter

Kyriazidou (1997) attrition model

Start-up Convert Observations

Retention 0.0024 (0.0014) 0.0001 (0.0001) 1,777,994

THE REVIEW OF ECONOMICS AND STATISTICS

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

THE REVIEW OF ECONOMICS AND STATISTICS

(1) (2) (3) (4) (5) (6) (7)

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

THE REVIEW OF ECONOMICS AND STATISTICS

Figure 7.Discipline and Attendance by before and after Charter Exit

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

THE REVIEW OF ECONOMICS AND STATISTICS

ACHIEVEMENT AND BEHAVIOR IN CHARTER SCHOOLS

You might also like